Exploiting Image Indexing Techniques in DCT Domain

This paper is concerned with the indexing and retrieval of images based on features extracted directly from JPEG discrete cosine transform (DCT) domain. We examine possible ways of manipulating DCT coefficients by standard image analysis approaches to describe image shape, texture, and color. Through Mandala transformation, our approach groups a subset of DCT coefficients to form ten blocks. Each block represents a particular frequency of the original image. We use two blocks to model rough object shape; nine blocks to describe subband properties; and one block to compute color distribution. As a result, the amount of data used for processing and analysis is reduced significantly. This can lead to simple yet efficient ways, of indexing and retrieval in a large scale image database. Experimental results show that it only takes approximately 6ms to index shape features, 5ms to index texture features, and 8ms to index color features from an image the size of 128 x 128 on a Sun Sparc Ultra-1 machine.